Nonintrusive speech quality estimation based on Perceptual Linear Prediction

Abstract

Objective measures of speech quality are attractive because they facilitate performance assessment of hearing aids (HAs) without the need for human listeners. Objective speech quality predictions are usually performed intrusively, wherein the “closeness” between the reference and HA output speech recordings is quantified. In this paper, we focus on nonintrusive estimation of HA speech quality based on Perceptual Linear Prediction (PLP) modeling approach. In PLP, perceptual phenomena such as non-uniform filter bank analysis and nonlinear mapping between sound intensity and its perceived loudness are incorporated into the linear prediction feature extraction process. In this work, PLP and PLP-based cepstral coefficients were computed from HA speech recordings and their statistical properties were utilized as features for speech quality estimation. A custom database of HA speech recordings obtained in different noisy and reverberant environments was used to investigate the predictive performance of these individual features. In addition, regression functions that linearly combined the features and mapped to the predicted quality scores, were derived and validated. Experimental results show that the proposed nonintrusive speech quality estimates correlate well with subjective ratings of speech quality by hearing impaired listeners.